Unable to retarget eyes simply by left eye or right eye.

[hoveychen]

In the current impl, although the detected eye ratios has separates values for left eye and right eye, when invoking the retargeting module, only the driving left eye ratio is used.

Ref code here: https://github.com/KwaiVGI/LivePortrait/blob/54e50986b232fc3f382f20924cdff675c0ce729d/src/live_portrait_wrapper.py#L308

Any plan to add an option to retarget eyes with separate left/right eyes in the driving video? IMO, we may invoke the retargeting module twice, each time with a separated eye. Then merge both delta 3DMM keypoints

[ak01user]

mark

[hoveychen]

Tried to modify the codes, and it works as expected.

[ak01user]

Tried to modify the codes, and it works as expected.

could you show your codes modified,please!

[hoveychen]

Sure. Since I don't have any docs about the meaning of the landmarks, I guess it would be something similar to https://github.com/anilbas/BFMLandmarks/blob/master/img/21.jpg

            combined_eye_ratio_tensor_left, combined_eye_ratio_tensor_right = (
                self.live_portrait_wrapper.calc_combined_eye_ratio_dual(
                    c_d_eyes_i, source_lmk
                )
            )
            # ∆_eyes,i = R_eyes(x_s; c_s,eyes, c_d,eyes,i)
            eyes_delta_left = self.live_portrait_wrapper.retarget_eye(
                x_s,
                combined_eye_ratio_tensor_left,
            )
            eyes_delta_right = self.live_portrait_wrapper.retarget_eye(
                x_s,
                combined_eye_ratio_tensor_right,
            )

            weights = (
                torch.tensor(
                    [
                        1,
                        1,
                        1,
                        0,
                        0,
                        0,
                        0.5,
                        0.5,
                        0.5,
                        0.5,
                        0.5,
                        1,
                        0.5,
                        0.5,
                        0.5,
                        0,
                        0.5,
                        0.5,
                        0.5,
                        0.5,
                        0.5,
                    ]
                )
                .view(21, 1)
                .to(device)
            )
            eyes_delta_left = eyes_delta_left.reshape(-1, 3)
            eyes_delta_right = eyes_delta_right.reshape(-1, 3)

            eyes_delta = eyes_delta_left * weights + eyes_delta_right * (1 - weights)

[ak01user]

Sure. Since I don't have any docs about the meaning of the landmarks, I guess it would be something similar to https://github.com/anilbas/BFMLandmarks/blob/master/img/21.jpg

            combined_eye_ratio_tensor_left, combined_eye_ratio_tensor_right = (
                self.live_portrait_wrapper.calc_combined_eye_ratio_dual(
                    c_d_eyes_i, source_lmk
                )
            )
            # ∆_eyes,i = R_eyes(x_s; c_s,eyes, c_d,eyes,i)
            eyes_delta_left = self.live_portrait_wrapper.retarget_eye(
                x_s,
                combined_eye_ratio_tensor_left,
            )
            eyes_delta_right = self.live_portrait_wrapper.retarget_eye(
                x_s,
                combined_eye_ratio_tensor_right,
            )

            weights = (
                torch.tensor(
                    [
                        1,
                        1,
                        1,
                        0,
                        0,
                        0,
                        0.5,
                        0.5,
                        0.5,
                        0.5,
                        0.5,
                        1,
                        0.5,
                        0.5,
                        0.5,
                        0,
                        0.5,
                        0.5,
                        0.5,
                        0.5,
                        0.5,
                    ]
                )
                .view(21, 1)
                .to(device)
            )
            eyes_delta_left = eyes_delta_left.reshape(-1, 3)
            eyes_delta_right = eyes_delta_right.reshape(-1, 3)

            eyes_delta = eyes_delta_left * weights + eyes_delta_right * (1 - weights)

It worked,very thanks.so we can try more things from editing 21 landmarks.that is much cool.

[ak01user]

hi,I want to edit param_weights,How do you know the value of param_weights? @hoveychen

[hoveychen]

Good question. As I mentioned above, I don't know the exactly mapping of the landmarks (which is then used to compose the mask weights). I just guess by trying to set different eye retargeting ratio. I assumed that the index 0-2 is the left eyebrow, 3-5 is the right eyebrow, 11 is the left eyeball, 15 is the right eyeball.

[hoveychen]

My previous comment was not correct. Check out this one.

11-13: right eye, 14-16 left eye, 12 and 15 are the eye balls.

[wangli000]

My previous comment was not correct. Check out this one.

11-13: right eye, 14-16 left eye, 12 and 15 are the eye balls.

hello, is this a visualization of the implicit keypoints in the paper？